RewardBench
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/allenai/reward-bench
下载链接
链接失效反馈官方服务：
资源简介：
该数据集名为RewardBench，是一个广泛应用的基准测试工具，用于评估奖励模型的能力及其安全性，主要通过成对的偏好任务来进行评估。它包含四个主要类别，旨在评估LLMs（大型语言模型）在聊天、难度较高的聊天、推理（数学加编程）以及安全性方面的特定期望能力。该数据集规模涵盖了23个独立的数据集，其任务在于评估奖励模型的能力。
This dataset, named RewardBench, is a widely adopted benchmark tool for evaluating the capabilities and safety of reward models, primarily via pairwise preference tasks. It includes four main categories, designed to assess the specific desired capabilities of LLMs (Large Language Models) in general chat, high-difficulty chat, reasoning (mathematics and programming), and safety. This benchmark covers 23 independent datasets, whose core task is to evaluate the capabilities of reward models.
提供机构：
Authors of the paper
